Anaphoric dependencies: How are they encoded?
Towards a derivation-based typology[1]
Eric Reuland
Utrecht institute of Linguistics OTS
1. Introduction
The typology of reflexives presented in Faltz (1977) had a major impact on the research in this field. But since Faltz's work, the study of reflexives and related issues has undergone significant developments. In addition to reflexives, reciprocals have recently become the subject of systematic studies as well. Rather than being taken for granted, the notion of binding itself has become the subject of investigation, and the distinction between anaphors and pronominals turned out to be less clear cut than previously assumed. In this contribution I will address these developments, focusing on the following three questions:
i. Can a typology of reciprocals be modeled on the typology developed by Faltz (1977) for reflexives? And how could such a typology be refined?
ii. What is (syntactic or semantic) 'binding'? Can reciprocity and reflexivity both be described in terms of that notion?
iii. Does the term 'anaphor' have any theoretical significance, or is it merely a convenient label used to refer to a specific class of pronominals or pronoun-like elements with certain ('defective') referential properties? Can elements like English 'each other' be said to be 'anaphors'?
I will start with the second question, and show that it has in fact more ramifications than one may have initially thought. Thus, a great deal of the paper will be devoted to a discussion of issues revolving around (ii). It will be shown that a typology of reflexives is needed that uses dimensions complementing Faltz's original typology. In short, whereas Faltz's typology is based on the morphological status of the reflexive elements, the origin of their composite parts, and the way they are morpho-syntactically connected, we will see that there are further dimensions, which reflect the role of reflexives in the derivation of the reflexive interpretation.
In the end, both (i) and (iii) will receive brief answers in terms of what we can conclude about (ii).
This article is organized as follows: In section 2. I discuss some basic theoretical and empirical issues in the theory of binding, and of the way binding relations are encoded in language. The chapter starts out with the canonical binding theory, followed by a discussion as to why it needs revision, and a brief overview of reflexivity theory. In view of the fact that anaphors are often analyzed as deficient in some respect, a discussion of the notion of under-specification is presented. Two fundamental why-questions are introduced and assessed: Why must reflexivity be licensed and why must anaphors be bound? The first question is answered, the second question is left for discussion in section 3. Finally the question is addressed what reflexive binding and reciprocal binding have in common. Section 3 focuses on two scenario's for deriving the binding requirement on complex anaphors, one for SELF-anaphors, the other for body-part reflexives. It is shown that under certain general empirical assumptions the binding requirements for SELF-anaphors and body-part reflexives can be derived without recourse to any statement that is specific to binding. Section 4, finally, provides a summary of the typology of reflexives and reciprocals, based on the results of the previous sections.
2. What is (syntactic or semantic) 'binding'? Can reciprocity and reflexivity both be described in terms of that notion?
As is well-known, languages show a variety of interpretive dependencies. (1) illustrates some of them, with the elements involved in the dependencies italicized:
(1) Dependencies
· Question formation (wh-movement)
o What do you think John saw _?
· Donkey anaphora
o Every man who owns a donkey beats it.
· Scopal dependencies
o Three men lifted a table.
Clearly not all of these dependencies are of the same sort, and they differ from canonical binding cases as in (2).
(2) Binding
o No one thought that he would have to leave.
o John hated himself.
Yet, the term "binding" is sometimes used in a very broad sense so as to include many types of dependencies. If the question of whether reciprocity and reflexivity can both be described in terms of binding is to have empirical content, the notion of binding should be made precise. Moreover, if we are to assess whether various dependencies in different languages do in fact instantiate binding, our definition of binding should also be independent of the ways in which languages encode it.
In accordance with standard practice I restrict discussion to binding relations between elements in "argument positions", positions that can be Case- or theta-marked (generally referred to as A-binding), and I will take standard reconstruction mechanisms for dislocated elements (involving A'-binding) for granted.[2] The canonical theory of A-binding (Chomsky 1981) is based on the following ingredients:
i. A division of argument types in:
A. Anaphors
B. Pronominals, and
C. R-expressions.
ii. The notion of an index to represent dependencies.
iii. C-command as a structural condition on binding.
iv. Syntactically defined locality conditions (captured by the notion of a governing category).
Binding is defined as follows:
v. a binds b iff
- a and b are co-indexed, and
- a c-commands b:
a is a sister of g containing b, as in the configuration a [g …b…]
The Binding conditions are formulated as:
A. An anaphor is bound in its governing category
B. A pronominal is free in its governing category
C. An R-expression is free
The notion of a governing category captures the locality effects which binding of pronominals and anaphors exhibits. It is defined as follows:
g is a governing category for a if and only if g is the minimal category containing a, a governor of a, and a SUBJECT (accessible to a)[3]
Although this canonical formulation of the binding theory captures a considerable range of facts, over the last decades it has become clear that it must be revised. Below, I summarize some of these reasons. After some preliminaries I will come back to these reasons in more detail.
2.1. Reasons for revision
A couple of reasons why the canonical binding theory had to be revised are given below:
· There are systems with more distinctions than just the distinction between anaphor and pronominal. For instance (limiting ourselves to a very small subset of cases to exemplify the point):
o Dutch has a 3-way system: pronominals such as hem 'him', simplex anaphors (henceforth, SE-anaphors) such as zich 'himself', complex anaphors (SELF-anaphors) such as zichzelf 'himself'.
o Icelandic, and Norwegian with the other mainland Scandinavian languages) have a 4-way system: Pronominals, SE-anaphors, SE-SELF and Pronominal-SELF.
· In addition to structural conditions, properties of predicates play a role in determining binding possibilities as well:
o English has John washed (no object) with a reflexive interpretation, but not *John hated
o Dutch has Jan waste zich (a SE-anaphor), but not *Jan haatte zich, etc.
· Under certain structurally defined conditions certain anaphoric forms need not be locally bound, or not even be bound at all:[4]
o Exempt himself in English
o Logophoric sig in Icelandic
· Certain languages allow locally bound pronominals
o him in Frisian: Jan waske him[5]
o 1st and 2nd person pronominals across the board: Ich wasche mich, jij wast je, nous nous lavons, etc.
These facts entail that it is impossible to provide an independent characterization of anaphors versus pronominals in terms of an intrinsic obligation to be locally bound or free. Features such as [+ anaphor] and [+ pronominal] (Chomsky 1981 and subsequent work) are not primitive lexical features.[6]
In order to capture these facts a modular approach to binding was developed (Reinhart and Reuland 1991, 1993, Reuland and Reinhart 1995, elaborated in Anagnostopoulou and Everaert 1999, and subsequently in Reuland 2001, Reinhart 2002, Reinhart and Siloni 2005, and others). Binding relations are licensed by intrinsic features of binder and bindee, together with the properties of the predicates they are arguments of. It would carry us beyond the scope of the present article to give a full overview. However, I will present a brief exposition of the main issues and their implications for typology.
2.2. A brief overview of issues in reflexivity
The typology of anaphoric expressions developed in Reinhart and Reuland (1993) and modified in Anagnostopoulou and Everaert (1999) is presented in Table 1.
Table 1 A typology of anaphoric expressions
SELF SE PRONOMINAL PRON/SELF
Reflexivizing function: + - - +
R(eferential independence): - - + +
SELF stands for elements like English himself, Dutch zichzelf, SE for Dutch zich, Icelandic sig, etc. PRONOMINAL for him, etc. and PRON/SELF for Greek o eaftos tou
· SELF-anaphors are reflexivizers: they license the reflexive interpretation of a predicate (as more precisely expressed in Conditions A and B below).[7]
· Pronominals and SE-anaphors (e.g. Dutch zich, Icelandic sig) do not by themselves license a reflexive interpretation of a predicate.
o Thus, if pronominals and SE-anaphors occur as 'bound' complements of reflexive predicates they do not mark these predicates as reflexive. They only reflect that these predicates are reflexive for an independent reason.
o This difference between SELF-anaphors and SE-anaphors is illustrated by the ill-formedness of Jan haatte zich 'John hated SE'. Zich does not license reflexivity here. In order for the reflexive interpretation to be licit, the complex anaphor zichzelf must be used, as in Jan haatte zichzelf 'John hated himself'.
o Hence, the technical notion of reflexive-marking must be distinguished from the notion of "marking" as it is found in much of the typological literature.
· Pronominals and SE-anaphors are alike in that both consist solely of phi-features: person, number, gender and a feature for a syntactic category; in addition they bear Case. They can be characterized as +/-R(eferential)
o The property +/-R has two faces: It is a semantic notion, but grounded in morpho-syntax:
§ Semantically +R can be characterized as standing for: capable of independent reference (SE-anaphors and SELF-anaphors cannot be used deictically);
§ Morphosyntactically +R-elements are all fully specified for phi-features and structural Case (Note this is a first approximation; for discussion of full versus underspecification, see below);
§ Syntax can only see the morpho-syntactic properties of pronouns.
o Pronominals are +R
o SE-anaphors (and SELF-anaphors) are –R.
· As Anagnostopoulou and Everaert (1999) show, the reflexivizing function and referential independence can be combined, as in the Greek element o eaftos tu that on the one hand is a reflexivizer, and on the other can occur in subject position without being bound.
The binding facts are captured by two conditions on predicates and a condition on A-chains, given in table 2, and briefly illustrated:
Table 2
Binding Conditions
Condition A: A reflexive-marked syntactic predicate is reflexive
§ Captures the following types of facts (among others):
o *I saw himself is ill-formed: himself is a syntactic argument of saw, hence forces the predicate to be reflexive, the feature mismatch between subject and object makes it impossible for this requirement to be satisfied.
o In Maxi expected the queen to invite Mary and himselfi for a drink the relevant argument of invite is Mary and himself, which properly contains himself; hence himself is exempt from condition A, and does not reflexivize the predicate. As a consequence, himself may receive a non-local antecedent. In related cases, depending on the structure himself need not have a linguistic antecedent at all (see for instance, Pollard and Sag 1992, 1994).[8]
Condition B: A reflexive semantic predicate is reflexive-marked.
§ Captures (among other things) that the collective we elected me is well-formed and the distributive we voted for me is awkward; the distributive reading yields a reflexive instantiation of the vote-for relation (a semantic predicate of the form x voted for x) which is not (and cannot be licensed). The same holds true for Johni hates Mary and himselfi/*himi. The reflexive instantiation of the hate-relation must be licensed, which requires himself instead of him. It also captures the contrast in the binding of him between Johni expected [[Mary and himi] to be in danger] and Johni persuaded [Mary and him*i] [PRO to leave], since John and him are coarguments of the same semantic predicate in the latter, but not in the former (ECM) case.
Condition on A-chains
A maximal A-chain (a1,..., an) contains exactly one link - a1 -which is +R
An NP is +R iff it carries a full specification for Φ-features and structural Case
Binding conditions A and B say nothing about the contrast between the pronominal and the SE-anaphor in (3):
(3) a. Jan waste zich/*hem.
b. Jan voelde [zich/*hem wegglijden].
This contrast follows from the chain condition. The chain condition captures that in Dutch, German, Icelandic, Mainland Scandinavian, and many other languages, 3rd person pronominals cannot be locally bound. In both (3a) and (3b) <Jan, zich> and <Jan, hem> are chains. Hem is fully specified for phi-features, namely 3rd person, masculine, singular, and it is in a position of structural Case. As a consequence, the chain <Jan, hem> violates the chain condition. Zich is only specified for 3rd person; it is incompatible with 1st or 2nd person antecedents, but fully compatible with feminine, masculine and neuter, and with singular and plural antecedents. Thus, zich is not fully specified for phi-features and the chain <Jan, zich> obeys the chain condition.
The chain condition facts can be typically assessed in configurations where conditions on reflexivity are satisfied as in *Jan waste hem with intrinsically reflexive wassen or *Jan voelde [ hem wegglijden] 'John felt SE slip away' where Jan and zich are not co-arguments. In the form in which it is presented here, the chain condition is essentially a descriptive generalization over representations. On a more general level, one would like to know why it holds. As a first step it is important to find out which factors enter into determining +/- R status. Hence we turn to some of the puzzling facts: locally bound pronominals, and set out to understand in what respects they differ from pronominals that cannot be locally bound.
2.2. 1. Locally bound pronominals
As noted above, in many languages, 1st and 2nd person pronominals can be locally bound, unlike 3rd person pronominals. Benveniste (1966) showed that their feature specifications differ in two respects: specification for person and specification for number. He argued that what is usually called "3rd person" is in fact absence of person (non-person)[9]. For his argument the number property is the key. Consider 3rd person/non-person elements. A plural noun such as dogs stands for a plurality of elements meeting the dog-criterion; a plural pronoun such as they can stand for a plurality of elements meeting the criteria for being a he, she or it. This is different for 1st and 2nd person. In 1st person, we is not marked for plural in that sense: we is not a plurality of I's. Rather it is inherently plural. In the 2nd person, plural you does not necessarily stand for a plurality of addressees (you can be talking to one person using plural you, including the people that are part of the addressee's "group"). So, the nature of "number" in 1st and 2nd person pronominals differs from that in 3rd person pronominals; only the latter show a grammatical number contrast, that is, changing the number only affects plurality versus singularity.[10] Therefore, if, for the number property, carrying a full specification for phi-features is narrowed down to the requirement of "being specified for grammatical number", local binding of 1st and 2nd person pronouns does not violate the chain condition.